fix(test): tolerate the prune race in picker alt-x branch assertions

[worktrunk-bot]

Problem

The nightly feature-powerset job failed on main when commands::picker::tests::test_apply_drops_unmorphable_kept_branch_row panicked with called Result::unwrap() on an Err value: git branch --list feature failed (exit 128) (failed run).

Root cause is a background-removal race in the test, not the picker. The picker's alt-x removal runs on a detached thread: stage_worktree_removal renames the worktree into the trash — so its directory vanishes — and then runs git worktree prune, which deletes the .git/worktrees/<id> admin dir. The three apply-based tests poll only on the worktree directory disappearing (which happens at the rename, before the prune) and then immediately query git branch --list in the main repo. git branch --list enumerates worktrees to mark checked-out branches, so a query that races the in-flight prune reads a half-deleted admin dir and fails with exit 128. The test .unwrap()ed that query and panicked.

The failure surfaced under the nightly job's serial, heavily-loaded run; the test predates this commit (from #3268) — this was a latent flake, not a regression.

Solution

Add a retry-tolerant await_branch_presence test helper that polls git branch --list <branch> until git succeeds and the branch reaches its expected steady state (a transient prune race surfaces as Err, never a successful empty read, so only a successful query is authoritative). Route all three background-removal tests through it:

• test_apply_morphs_unmerged_worktree_to_branch_row — branch retained
• test_apply_drops_unmorphable_kept_branch_row — branch retained (the one that failed)
• test_apply_drops_integrated_worktree_row — branch deleted

The synchronous do_removal tests and the branch-only keep test are unaffected — they run no concurrent prune — so they're left alone. This is a test-synchronization fix; picker behavior is unchanged.

The second failing job in the same run, nix-flake (test_remove_reap_kills_process needing lsof), is already addressed by #3410 and #3406, so it's out of scope here.

Testing

• cargo test --no-default-features --features cli,git-wt,shell-integration-tests,syntax-highlighting --bin wt for the three tests — pass.
• Stress: ran all three under --test-threads=8 for 40 iterations — 0 failures.
• cargo fmt --check and cargo clippy --bin wt --tests (same feature set) — clean.

---

Automated fix for failed run

[worktrunk-bot]

The flake diagnosis and fix are sound — I traced all three call sites and the list.is_empty() != expect_present logic is correct for both the retained (true) and deleted (false) cases, and tolerating the transient exit 128 (retry on Err, only a successful query is authoritative) is exactly right for the prune race.

One non-blocking reuse note below: the hand-rolled poll loop duplicates an existing shared test helper. Not a correctness concern — flagging per the repo's "check before hand-rolling a utility" convention.